Arbor: Hypothesis-Tree AI Optimization Framework Beats Claude Code & Codex by 2.5x [2026]

Arbor is a long-horizon autonomous optimization framework developed by Renmin University of China and Microsoft Research. At its core is Hypothesis Tree Refinement (HTR), a persistent hypothesis tree structure that enables AI agents to conduct research and optimization autonomously. It achieves over 2.5x average held-out performance gains compared to Claude Code and Codex under the same compute budget.

📑Table of Contents

1. What is Arbor? The Basics of Hypothesis Tree Refinement (HTR)
2. Coordinator and Executor Role Separation and Worktree Utilization
3. Accumulating and Propagating Failures with the Persistent Hypothesis Tree
4. Benchmark Results: Comparison with Claude Code/Codex and 2.5x Gains
5. BrowseComp Task Improvement Example (45.33% → 67.67%)
6. GitHub Repository and arXiv Paper Public Content
7. Practical Implications for AI Agent Developers
8. Frequently Asked Questions (FAQ)
9. Comparison Table
10. Summary

---

What is Arbor? The Basics of Hypothesis Tree Refinement (HTR)

The heart of Arbor is the Hypothesis Tree Refinement (HTR) method. It persistently manages hypotheses in a tree structure, accumulating and refining them with experimental results and evidence. Rather than one-off task execution, it updates and propagates hypotheses over the long term to enhance AI agent autonomy. The arXiv paper (https://arxiv.org/abs/2606.11926) demonstrates its effectiveness on Autonomous Optimization tasks such as model training and data synthesis.

HTR’s advantage lies in linking hypotheses with artifacts and distilling insights for future use. This overcomes limitations of existing tools like context loss and reward hacking.

---

Coordinator and Executor Role Separation and Worktree Utilization

Arbor clearly separates two roles: Coordinator and Executor. The Coordinator handles long-term strategy and manages the overall hypothesis tree. The Executor runs experiments in isolated worktree environments and safely feeds back results. This separation prevents experimental failures from affecting the entire system, enabling stable operation.

The GitHub repository (https://github.com/RUC-NLPIR/Arbor) publishes examples using worktrees, and setup procedures can be verified there. The project page (https://ruc-nlpir.github.io/Arbor/) is also useful.

---

Accumulating and Propagating Failures with the Persistent Hypothesis Tree

Arbor’s strength is treating failures not as mere errors to discard but as constraints to accumulate and propagate. The hypothesis tree is persistent, so past failures and evidence influence future decisions. This allows AI agents to adjust strategies more intelligently and achieve long-term optimization.

VentureBeat coverage also highlights this mechanism as the reason it surpasses Claude Code and Codex. Utilizing failures differentiates it from conventional approaches where insights dissipate easily.

---

Benchmark Results: Comparison with Claude Code/Codex and 2.5x Gains

Benchmarks show Arbor’s clear superiority over Claude Code and Codex. Under the same compute budget, it recorded average held-out performance gains of over 2.5x. On MLE-Bench Lite, it achieved 86.36% Any Medal with GPT-5.5, the highest among comparators.

See the comparison table in the configuration section. Arbor excels in long-term strategy management and failure utilization.

---

BrowseComp Task Improvement Example (45.33% → 67.67%)

On the BrowseComp task, a significant improvement from 45.33% baseline to 67.67% was confirmed. This task involves complex optimization with web browsing, where HTR’s hypothesis management proved effective. It is backed by detailed experimental results in the arXiv paper.

---

GitHub Repository and arXiv Paper Public Content

Arbor’s code is public on GitHub (https://github.com/RUC-NLPIR/Arbor), and the arXiv paper (https://arxiv.org/abs/2606.11926) details the theoretical background and evaluation results. The project page (https://ruc-nlpir.github.io/Arbor/) provides demos and additional materials. VentureBeat (https://venturebeat.com/orchestration/new-ai-optimization-framework-beats-claude-code-and-codex-by-2-5x-on-the-same-compute-budget) also reports from a practical perspective.

---

Practical Implications for AI Agent Developers

For engineers developing AI agents, Arbor provides valuable insights for next-generation workflow design. Introducing Coordinator/Executor separation and HTR improves stability and performance on long-horizon tasks. We recommend cloning the GitHub repository and trying it in a local environment first.

---

Frequently Asked Questions (FAQ)

1. Is Arbor a replacement for Claude Code or Codex?
   No, Arbor is a framework that complements and extends existing tools. By combining Coordinator and Executor, it further enhances the performance of Claude Code and Codex.

2. How does Hypothesis Tree Refinement (HTR) manage hypotheses?
   It persistently holds hypotheses in a tree structure, linking experimental results and evidence to refine them. Failures are also accumulated as constraints.

3. How do Coordinator and Executor collaborate?
   The Coordinator formulates overall strategy, while the Executor executes experiments in isolated worktrees. Results are fed back into the tree for collaboration.

4. On which tasks was the 2.5x performance gain confirmed?
   Average gains of over 2.5x across Autonomous Optimization tasks in general. Specific numbers are shown on MLE-Bench Lite and BrowseComp.

5. What does the 86.36% Any Medal on MLE-Bench Lite mean?
   It indicates the any-medal acquisition rate, representing top-tier results among comparators. It was achieved in combination with GPT-5.5.

---

Comparison Table

Item	Claude Code / Codex	Arbor (HTR)
Long-term strategy management	Limited	Persistent tree by Coordinator
Failure utilization	Prone to dissipation	Accumulated and propagated as constraints
Held-out performance gain	Baseline	Average 2.5x or more
BrowseComp improvement example	–	45.33% → 67.67%
Isolated execution environment	Standard	Executor worktree separation

Source: arXiv:2606.11926 (June 2026), VentureBeat coverage, official GitHub and project page (as of June 2026)

---

Related articles:

• Boltz Bio、BoltzMol-1 / BoltzProt-1 をリリース — 創薬向け新モデル、Claude Code / Codex 統合対応
• MiniMax M3 リリース — 初のオープンウェイト frontier モデル（1M コンテキスト + ネイティブ multimodal + 59% SWE-Bench）
• Databricks、AI Agent向けメタハーネス「Omnigent」をオープンソース公開 — Claude Code / Codex横断でmulti-agent制御

Summary

Arbor opens new possibilities for AI agent development as a long-horizon autonomous optimization framework leveraging Hypothesis Tree Refinement. The separation of Coordinator and Executor and persistent failure management deliver performance far exceeding Claude Code and Codex. Check the details via GitHub and arXiv, and try incorporating it into your projects.